The human cell is wildly complex. Can AI decode it? | Silvana Konermann

In this TED Talk, Silvana Konermann, a bioengineer and neuroscientist at the ARC Institute and 2025 Audacious Project grant recipient, discusses her vision for using AI to decode the complexity of human cells and accelerate treatment of complex diseases like Alzheimer's, heart disease, and cancer. She explains that 'complex diseases' — as a technical term — are those driven by multiple interacting genetic and environmental risk factors unique to each patient, which is why they have resisted medical breakthroughs despite decades of research.

Konermann identifies three converging technological advances making this vision newly feasible: single-cell RNA sequencing (which captures gene expression snapshots one cell at a time), CRISPR-based gene editing (which allows precise, targeted perturbation of individual genes), and modern AI (which can learn patterns from vast biological datasets much as large language models learn from human text). She draws an explicit analogy between AI learning human language and AI learning the 'language' of cells via RNA expression data, noting that unlike human language, biological language was not created by humans and is therefore largely impenetrable to human intuition — making AI particularly well-suited to decode it.

To train her model, Konermann's team plans to conduct at least one billion physical cellular experiments over four years, perturbing one gene at a time via CRISPR and measuring the cellular response via RNA sequencing. Her team has already completed approximately 60 million such experiments. The resulting model — a 'universal virtual cell' — is intended to generalize across cell types and disease states it has never been trained on, enabling researchers to input a diseased cell state and receive predictions about which interventions would most likely convert it to a healthy state.

The ARC Institute released a first-generation model eight months prior to the talk. Konermann openly acknowledges it is only about 20% accurate and still far from clinically useful, but describes it as the current state of the art. She plans to release an interface tool later in the year for broader researcher use, with ongoing annual 'virtual cell challenges' to engage the global research community — over 1,000 teams participated in the first challenge. The model and tools are being made openly available rather than licensed exclusively to commercial entities.

On biosafety concerns, Konermann argues the tool's risk is limited because it is focused on human cells, not pathogens, and shifting human cell states is difficult to weaponize. She also notes that the model could accelerate pandemic response by identifying how viruses target cellular pathways. She estimates that within four to five years, models accurate enough to meaningfully change how biology and drug discovery are conducted will be available, shifting the field away from one-hypothesis-at-a-time drug development toward comprehensive, data-driven target identification.